replace the leak check with universes, take 2
This PR is an attempt to revive the "universe-based region check", which is an important step towards lazy normalization. Unlike before, we also modify the definition of `'empty` so that it is indexed by a universe. This sidesteps some of the surprising effects we saw before -- at the core, we no longer think that `exists<'a> { forall<'b> { 'b: 'a } }` is solveable. The new region lattice looks like this:
```
static ----------+-----...------+ (greatest)
| | |
early-bound and | |
free regions | |
| | |
scope regions | |
| | |
empty(root) placeholder(U1) |
| / |
| / placeholder(Un)
empty(U1) -- /
| /
... /
| /
empty(Un) -------- (smallest)
```
This PR has three effects:
* It changes a fair number of error messages, I think for the better.
* It fixes a number of bugs. The old algorithm was too conservative and caused us to reject legal subtypings.
* It also causes two regressions (things that used to compile, but now do not).
* `coherence-subtyping.rs` gets an additional error. This is expected.
* `issue-57639.rs` regresses as before, for the reasons covered in #57639.
Both of the regressions stem from the same underlying property: without the leak check, the instantaneous "subtype" check is not able to tell whether higher-ranked subtyping will succeed or not. In both cases, we might be able to fix the problem by doing a 'leak-check like change' at some later point (e.g., as part of coherence).
This is a draft PR because:
* I didn't finish ripping out the leak-check completely.
* We might want to consider a crater run before landing this.
* We might want some kind of design meeting to cover the overall strategy.
* I just remembered I never finished 100% integrating this into the canonicalization code.
* I should also review what happens in NLL region checking -- it probably still has a notion of bottom (empty set).
r? @matthewjasper
Add GitHub issue templates
This PR adds GitHub's issue templates to the repository. Adding templates for two of the most common issues we create on the repository. We could add more specific templates (e.g. ICEs) depending on the response to initial templates. I've included a screenshot of what it looks like, and people can also try out the UI and specific templates, by going to [`XAMPPRocky/rust`](https://github.com/XAMPPRocky/rust/issues/new/choose).
<img width="1115" alt="Screenshot 2020-01-21 at 13 57 46" src="https://user-images.githubusercontent.com/4464295/72807027-c51baa00-3c56-11ea-8a4c-98238489b345.png">
Selectively disable sanitizer instrumentation
Add `no_sanitize` attribute that allows to opt out from sanitizer
instrumentation in an annotated function.
BtreeMap range_search spruced up
#39457 created a lower level entry point for `range_search` to operate on, but it's really not hard to move it up a level of abstraction, making it somewhat shorter and reusing existing unsafe code (`new_edge` is unsafe although it is currently not tagged as such).
Benchmark added. Comparison says there's no real difference:
```
>cargo benchcmp old3.txt new3.txt --threshold 5
name old3.txt ns/iter new3.txt ns/iter diff ns/iter diff % speedup
btree::map::find_seq_100 19 21 2 10.53% x 0.90
btree::map::range_excluded_unbounded 3,117 2,838 -279 -8.95% x 1.10
btree::map::range_included_unbounded 1,768 1,871 103 5.83% x 0.94
btree::set::intersection_10k_neg_vs_10k_pos 35 37 2 5.71% x 0.95
btree::set::intersection_staggered_100_vs_10k 2,488 2,314 -174 -6.99% x 1.08
btree::set::is_subset_10k_vs_100 3 2 -1 -33.33% x 1.50
```
r? @Mark-Simulacrum
error code examples: replace some more ignore with compile_fail
Now that #68664 has been merged and `compile_fail` attempts a full build rather than `--emit=metadata`, these errors should be caught by `compile_fail` and do not need to be ignored.
Mark fn map_or() as eagerly evaluated.
In the docs for option.rs and result.rs, it is noted for all *_or()
functions that they are eagerly evaluated, except for the map_or()
function.
This commit adds this missing documentation to the two files.
Closes#68866
implement proper linkchecker hardening
r? @JohnTitor
This implements proper linkcheck filtering... we might need to fiddle with a bit to adjust what is or isn't filtered, but this seems to work reasonable locally.
Rename -Zexternal-macro-backtrace to -Zmacro-backtrace and clean up implementation.
This is my attempt at dealing with https://github.com/rust-lang/rust/pull/66364#issuecomment-565517232, although I'm not sure it's the least disruptive one.
The behavior of `-Zexternal-macro-backtrace` was already to enable full macro backtraces for *all* macros, the only part of it that was specific to cross-crate macros was showing this when *not used*:
```
note: this error originates in a macro outside of the current crate
(in Nightly builds, run with -Z external-macro-backtrace for more info)
```
After this PR:
* the flag is renamed to `-Zmacro-backtrace`
* do we need to have a deprecation period? cc @rust-lang/compiler
* the message informing you about the flag is always shown when an expansion of a bang macro/attribute/derive is involved, not just cross-crate ones
* this accounts for most of the changes in tests
* we could perhaps only show it for the bang macro case? feels odd for derives
* `fix_multispans_in_std_macros` is split into `fix_multispans_in_extern_macros` and `render_multispans_macro_backtrace`
* this roughly reverts the non-behavioral parts of #46605, which combined the two functionalities
* not sure where the old `std_macros` name came from, perhaps the `<std macros>` synthetic "file"? even then, odd that `std` specifically was mentioned
* `render_multispan_macro_backtrace`, by default (i.e. without `-Zmacro-backtrace`), hides the `in this macro invocation` label specifically to avoid redundancy in the diagnostic
* that is, showing the macro use site is only useful when the diagnostic is inside the macro definition and the user can't otherwise tell which use site it applies to, not when the diagnostic is at/inside the use site already (which would make the label redundant)
* before, it was only checking for the situation in which a cross-crate macro *definition* span would be replaced with the invocation span, which both made the connection to redundancy unobvious, and didn't help with other redundancy (e.g. when the diagnostic was pointing to an argument inside the macro invocation)
* this accounts for the remaining test changes, which I've first noticed in https://github.com/rust-lang/rust/pull/66364#discussion_r356135967 but only later understood as part of this PR (hence the "redundancy" descriptions)
This PR is not needed for #66364, but it would help, as after this PR there's only one `.span_to_filename(...).is_macros()` check (i.e. for `<... macros>` synthetic "files") left in `rustc_errors`, and it's much more self-contained.
r? @petrochenkov
We now make `'empty` indexed by a universe index, resulting
in a region lattice like this:
```
static ----------+-----...------+ (greatest)
| | |
early-bound and | |
free regions | |
| | |
scope regions | |
| | |
empty(root) placeholder(U1) |
| / |
| / placeholder(Un)
empty(U1) -- /
| /
... /
| /
empty(Un) -------- (smallest)
```
Therefore, `exists<A> { forall<B> { B: A } }` is now unprovable,
because A must be at least Empty(U1) and B is placeholder(U2), and hence
the two regions are unrelated.
We currently have a kind of arbitrary check for `Verify` conditions
which says that if the "test region" is `'empty`, then the check
passes. This was added to fix#42467 -- it happens to be correct for
the purposes that we use verify bounds for, but it doesn't feel
generally correct. Replace with a more principled test.
In the docs for option.rs and result.rs, it is noted for all *_or()
functions that they are eagerly evaluated, except for the map_or()
function.
This commit adds this missing documentation to the two files.
stop using BytePos for computing spans in librustc_parse/parser/mod.rs
Computing spans using logic such as `self.token.span.lo() + BytePos(1)` can cause internal compiler errors like #68730 when non-ascii characters are given as input.
#68735 partially addressed this problem, but only for one case. Moreover, its usage of `next_point()` does not actually align with what `bump_with()` expects. For example, given the token `>>=`, we should pass the span consisting of the final two characters `>=`, but `next_point()` advances the span beyond the end of the `=`.
This pull request instead computes the start of the new span by doing `start_point(self.token.span).hi()`. This matches `self.token.span.lo() + BytePos(1)` in the common case where the characters are ascii, and it gracefully handles multibyte characters.
Fixes#68783.
bors
